Designated commodity / service selection support system
The designated goods and services selection support system addresses the challenge of keyword-based inaccuracies by using a vector database and language model to provide accurate, professional results for trademark registration, reducing the risk of rejection.
Patent Information
- Application Number
- JP2025244854
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-12-10
- Publication Date
- 2026-03-04
AI Technical Summary
Conventional databases and AI systems struggle with keyword-based searches for designated goods and services in trademark registration, leading to incorrect or incomplete results due to users' lack of familiarity with trademark practice, which can result in inappropriate descriptions and potential rejection.
A designated goods and services selection support system that uses a vector database and a large-scale language model to assist users in selecting appropriate terms by performing semantic vector searches and providing exact matches based on input terms, incorporating analysis and output control rules to ensure accuracy.
Enables users unfamiliar with trademark practice to easily select appropriate designated goods and services with a low risk of rejection by ensuring semantic accuracy and preventing hallucinations, thus adhering to professional criteria.
Smart Images
Figure 2026035898000001_ABST
Abstract
Description
[Technical Field]
[0001] The present invention relates to a designated product / service selection support system. [Background technology]
[0002] Conventionally, databases such as J-PlatPat provided by the National Center for Industrial Property Information and Training (INPIT) have been used to research the designated goods and designated services used in trademark registration applications (see Non-Patent Document 1). By searching these databases using product names or service names as keywords, it is possible to find out the designated goods and services, classifications, and similar group codes of trademarks actually adopted by the Japan Patent Office.
[0003] It is also possible to ask a large-scale language model such as ChatGPT about designated goods and services and receive an answer (see Non-Patent Document 2). [Prior art documents] [Patent documents]
[0004] [Non-Patent Document 1] Product / service name search, [Search date August 25, 2025], Internet:<URL:https: / / www.j-platpat.inpit.go.jp / t1201> [Non-patent document 2] ChatGPT, [Retrieved August 25, 2025], Internet:<URL:https: / / chatgpt.com / > Summary of the Invention [Problem to be solved by the invention]
[0005] However, conventional databases are based on keyword searches. As a result, people who are unfamiliar with trademark practice face the challenge of being unable to recall the appropriate official name to use as a keyword to search for, making it difficult to find the information they are looking for.
[0006] For example, Figure 9 shows the results of a product / service name search using the keyword "judicial scrivener" on J-PlatPat. As you can see, product and service names containing the keyword "judicial scrivener" appear, but these are not the services provided by judicial scriveners that users actually wanted to know about. If you simply list this as the designated service and apply for it, it will not become a trademark right that is truly necessary.
[0007] Figure 10 shows the results of searching for product and service names using the keyword "rental" on J-PlatPat. Here too, only one product or service name containing "rental" was found, but even though there are many other rental businesses, the search did not yield any other hits, and users were unable to get any hints as to the correct product or service name.
[0008] Furthermore, if the description of the designated goods and services in a trademark registration application is inappropriate, that in itself may be grounds for refusal. Those unfamiliar with trademark practice find it difficult to select the appropriate expressions for designated goods and services, and as a result may unintentionally make inappropriate descriptions, which could result in a more serious issue of risking rejection.
[0009] If you can't think of an appropriate way to describe the designated goods and services, you can ask a generation AI like ChatGPT. Figure 11 shows the results of asking GPT-5 about the specified product for the item "underwear." It is correct that "underwear" and "socks" are in Class 25. On the other hand, GPT-5 output that the similar group code for both "underwear" and "socks" was 25A01.
[0010] However, the correct similar group code for "underwear" is 17A02, and the correct similar group code for "socks" is 17A04. This is hallucination (incorrect responses that are not based on facts), which is often cited as a problem when using generative AI. Moreover, ChatGPT not only got the similarity group code wrong, but also included a hallucination in the explanation for socks, emphasizing that "they are often managed with the same similarity group code as underwear."
[0011] Hallucination in similar group codes can lead to serious errors in searching for prior registered trademarks, resulting in reasons for refusal. Hallucination can also occur in the description of the designated goods and designated services themselves, and since such designated goods and designated services are essentially no different from those written by someone unfamiliar with trademark practice, writing them as they are can be a reason for refusal.
[0012] Therefore, the object of the present invention is to provide a designated goods and services selection support system that enables even a person unfamiliar with trademark practice to easily select appropriate designated goods and designated services using everyday language with a low risk of rejection. [Means for solving the problem]
[0013] In order to achieve the above object, the designated goods and services selection support system according to claim 1 of the present invention comprises: A system for supporting the selection of designated goods or designated services, A management server (10) for providing a designated product / service selection support service; a user terminal (1) configured to be able to communicate with the management server (10) via a network (100); The management server (10) a control unit (11) that executes processing related to the designated product / service selection assistance service; a storage unit (12) having a vector database in which semantic vectors based on the designated goods and designated services of trademarks actually adopted by the Patent Office are stored in advance; The control unit (11) Acquires input terms indicating products or services input from the user terminal (1), A vector search is performed on the vector database using the input term, and a plurality of designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Patent Office are extracted; Sending the extracted designated goods or designated services and the input terms, together with a prompt, to a large-scale language model (2) accessible via a network (100); Execute a process of outputting the analysis results returned from the large-scale language model (2) to the user terminal (1); The prompt may include: (a) an analysis rule for identifying, from among the extracted designated goods or designated services, those that match the meaning indicated by the input term; (b) an output control rule for outputting the specified designated goods or designated services exactly as they are stored in the vector database, without any word-for-word changes.
[0014] The designated goods and services selection support system according to claim 2 comprises: The original data for obtaining the semantic vector is compiled as tabular data, the semantic vector is obtained by vectorizing the tabular data row by row, In the vector database, each of the semantic vectors is treated as one chunk.
[0015] The designated goods and services selection support system according to claim 3 comprises: The original data for obtaining the semantic vector is compiled as tabular data, The tabular data includes the English names of the designated goods and designated services of the trademarks actually adopted by the Patent Office, The semantic vector is characterized in that it is made more accurate by using the English notation.
[0016] The designated goods and services selection support system according to claim 4 comprises: The analysis rule is: The names of the designated goods and services; and explanation data that explains the meaning of the name.
[0017] Here, the symbols in parentheses indicate corresponding elements or items shown in the drawings and in the detailed description to be described later. [Effects of the Invention]
[0018] According to the present invention, a vector search is performed on a vector database using an input term, multiple designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Patent Office are extracted, the extracted designated goods or designated services and the input term are sent, along with a prompt, to a large-scale language model that can be accessed via a network, and the analysis results returned from the large-scale language model are output to the user's terminal.As a result, even those who are unfamiliar with trademark practice can easily select appropriate designated goods or designated services using everyday language that have a low risk of rejection. In other words, even someone unfamiliar with trademark practice can extract designated goods or designated services from the designated goods or designated services actually adopted by the Patent Office, so the indication of the designated goods or designated services will not be unclear, and a reason for refusal (Article 6, Paragraph 1 of the Trademark Act) will not be issued for this reason.
[0019] In particular, because it performs vector searches, even those unfamiliar with trademark practice can arrive at highly relevant candidates for designated goods and services simply by searching using everyday words that come to mind. In addition, the prompt includes analytical rules for identifying the extracted designated goods or designated services that match the meaning indicated by the input term, so that it is possible to derive designated goods or designated services that are deemed appropriate from the vector search candidates in line with the professional criteria of trademark practice and the context.
[0020] Furthermore, the prompt includes output control rules for outputting the identified designated goods or designated services exactly as they are stored in the vector database, without any word-for-word changes, thereby ensuring that the output is an exact match with the designated goods and designated services of the trademark actually adopted by the Patent Office, thereby preventing hallucination.
[0021] Furthermore, according to the present invention, the original data for obtaining semantic vectors is compiled as tabular data, and the semantic vectors are obtained by vectorizing the tabular data row by row. Since each semantic vector is treated as a single chunk in the vector database, it is possible to faithfully extract only relevant items from the normalized list. In other words, hallucination is extremely unlikely to occur, so even if a person unfamiliar with trademark practice enters a term, it is possible to select the appropriate designated goods or designated services.
[0022] Furthermore, according to the present invention, the designated goods and designated services of trademarks actually adopted by the Patent Office are in tabular form, and the tabular data includes the English notation of the designated goods and designated services of trademarks actually adopted by the Patent Office, and the semantic vectors are made more accurate by the English notation, thereby improving search accuracy. In other words, using information from English spelling improves embedding, reducing word sense ambiguity, which in turn improves the accuracy of semantic features in vector space and improves the accuracy of search results. This allows for more appropriate selection of designated goods or designated services.
[0023] Furthermore, according to the present invention, the analysis rules are based on the names of the designated goods and designated services and explanatory data that explains the meaning of the names, so that the designated goods and designated services can be more appropriately derived from the candidates searched by vector.
[0024] Furthermore, the above-mentioned Non-Patent Documents 1 and 2 do not at all mention the point that the designated goods and services selection support system and designated goods and services selection support program of the present invention perform a vector search on a vector database to extract designated goods or designated services that have actually been adopted by the Patent Office. [Brief explanation of the drawings]
[0025] [Figure 1] 1 is a block diagram showing a designated product / service selection support system according to a first embodiment of the present invention; [Figure 2] FIG. 1 is a diagram showing an example of tabular data of designated goods and designated services of trademarks actually adopted by the Japan Patent Office, which is the source data of semantic vectors. [Figure 3] FIG. 2 is a sequence diagram showing the processing of a program in the designated product / service selection support system according to the first embodiment of the present invention. [Figure 4] FIG. 10 is a diagram showing search results for "judicial scrivener" in the designated product / service selection support system according to the first embodiment of the present invention. [Figure 5] FIG. 10 is a diagram showing search results for "rental" in the designated product / service selection support system according to the first embodiment of the present invention. [Figure 6] FIG. 10 is a sequence diagram showing the processing of a program in the designated product / service selection support system according to the second embodiment of the present invention. [Figure 7] FIG. 11 is a sequence diagram showing the processing of a program in the designated product / service selection supporting system according to the third embodiment of the present invention. [Figure 8] FIG. 1 is a diagram showing an example of tabular data of designated goods and designated services of trademarks actually adopted by the Japan Patent Office, which is the source data of semantic vectors. [Figure 9]FIG. 10 is a diagram showing the search results for "judicial scrivener" in a conventional product / service name search on J-PlatPat. [Figure 10] FIG. 10 is a diagram showing the search results for "rental" in a conventional product / service name search on J-PlatPat. [Figure 11] This figure shows the answers given when the conventional ChatGPT (GPT-5) is asked what the designated product is for "underwear." DETAILED DESCRIPTION OF THE INVENTION
[0026] (First embodiment) A designated product / service selection support system according to a first embodiment of the present invention will be described with reference to FIGS. This designated product / service selection support system includes a management server 10 and a user terminal 1, which are connected to each other via a network 100 such as the Internet so that they can communicate with each other. The designated product / service selection support system is also connected to a large-scale language model 2 via a network 100. Here, the large-scale language model 2 is a form of accessing a service (large-scale language model 2 server) implemented on a server on the cloud.
[0027] The management server 10 is for providing a designated product / service selection support service and executes the core processing of the present invention. The management server 10 includes a control unit 11, a storage unit 12, a communication interface, and the like.
[0028] The user terminal 1 is a device that allows a user to input input terms and view analysis results, and may be a PC, smartphone, tablet, or the like.
[0029] A large-scale language model2 is a machine learning model that performs sentence generation and question answering by learning from large amounts of natural language data. Examples include OpenAI's GPT series and Google's Gemini series.
[0030] Next, the management server 10 will be described in detail. The control unit 11 executes the programs loaded in the storage unit 12, thereby functioning as various processing units that execute processes related to the designated product / service selection support service.
[0031] The storage unit 12 has a vector database in which the designated goods and designated services of trademarks actually adopted by the Patent Office are stored in advance as semantic vectors. The "designated goods and designated services of trademarks actually adopted by the Japan Patent Office" includes five data types found in J-PlatPat's Goods and Service Name Search (https: / / www.j-platpat.inpit.go.jp / t1201): "Examination Guidelines for Similar Goods and Services," "Nice Classification," "TM5 ID List," "Goods and Service Names Adopted in Examination," and "WIPO Madrid Goods and Services Manager."
[0032] Here, a semantic vector is a numerical vector that is created by converting text data (e.g., a description of a trademark's designated goods or services) into a numerical vector using natural language processing technology. A semantic vector represents the semantic features of a sentence in a high-dimensional space, making it possible to numerically calculate the semantic similarity between sentences. A vector database is a database that stores semantic vectors and has the function of searching based on the similarity with an input vector.
[0033] As shown in Figure 2, the original data for obtaining the semantic vectors of these "designated goods and designated services of trademarks actually adopted by the Patent Office" is compiled in tabular form as shown in Figure 2. In this tabular data, Japanese product and service names are associated with attributes such as classification, data type, English product and service names, and similar group codes.
[0034] The semantic vectors are obtained by vectorizing this tabular data row by row as a preprocessing step, and each semantic vector is treated as a single chunk in the vector database. A chunk is a unit obtained by dividing long data into pieces of data suitable for searching or processing. In this embodiment, one chunk corresponds to the data divided into each row of the tabular data of the designated goods or designated services of a trademark.
[0035] By using (or combining) English notation to generate semantic vectors, it is possible to reflect semantic features based on multiple languages and improve the accuracy of vector representations.
[0036] The functions executed by the control unit 11 of the designated product / service selection support system configured as above will be described below. By executing the program, the control unit 11 executes a series of processes, which will be described later, such as obtaining input terms, executing a vector search, linking with the large-scale language model 2, and outputting the results. Next, with reference to the flowchart in FIG. 3, a specific processing flow from when input terms are sent from the user terminal 1 to when the results are output will be described.
[0037] First, the service provider compiles the designated goods and designated services of trademarks that have actually been adopted by the Patent Office into tabular data, as shown in Figure 2. As a pre-processing step, the control unit 11 inputs the tabular data into an embedding model using a database construction program. The embedding model comprehensively interprets various information about each designated product or service, such as its name and English spelling, and generates a unique numerical coordinate that expresses its meaning, i.e., a semantic vector (Step S100 (hereafter, the word "Step" will be omitted)). This process makes it possible to grasp not only the superficial correspondence of words, but also their conceptual closeness.
[0038] The constructed vector database stores the generated semantic vector for each designated product or service, in association with attribute information such as classification, data type, English notation, and similar group code. Since S100 is a pre-processing step, it does not need to be done every time, and it is sufficient to perform it when the designated goods and designated services are revised, for example.
[0039] Next, when the user inputs an input term indicating a product or service, such as "judicial scrivener," from the user terminal 1, the management server 10 acquires this (step S101). The input terms do not necessarily represent goods or services known to trademark practitioners, but may be everyday words for the goods or services.
[0040] The control unit 11 converts the received input terms into query vectors using the embedding model (S102).
[0041] Next, the control unit 11 executes a vector search in the vector database using the query vector derived from the input term, thereby extracting multiple designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Patent Office (S103).
[0042] Next, the control unit 11 transmits the extracted designated goods or designated services as a candidate group (first candidate group) together with the input term and a prompt to the API of the large-scale language model 2 (S104). The prompt includes an analysis rule for identifying, from among the extracted designated goods or designated services, those that match the meaning indicated by the input term, and an output control rule for outputting the identified designated goods or designated services word for word as they are stored in the vector database.
[0043] The analysis rules incorporate the following decision logic: First, the user is asked to define the "core value" provided by the terms they input (for example, for "judicial scrivener," the value is "a professional who prepares and acts on behalf of others regarding registration"). Next, based on the defined value, the designated goods and designated services that directly represent the business are selected from the extracted list of designated goods and designated services. At that time, the applicant is asked to confirm whether the selected designated goods and designated services are consistent with the contents of the "Explanation of Classification of Goods and Services," which is explanatory data that explains the meaning of goods and services.
[0044] Next, in addition to the direct products and services mentioned above, we will expand the business possibilities and have you select designated products and services that correspond to each upstream and downstream development. This will allow us to propose products and services for related businesses in addition to the direct products and services.
[0045] Then, for each selected candidate, the system deeply analyzes the reasons why it matches the user's input terms, and generates a professional yet concise reason that is easy for the user to understand based on its relevance to the input terms. The above is an example of an analysis rule.
[0046] The output control rule is to output the identified designated goods or designated services word for word as they are stored in the vector database, without any changes, but the value of a large-scale language model 2 is usually its ability to "generate intelligent, natural sentences." However, the output control rule for the prompts in the designated goods and services selection support system according to this embodiment intentionally confines this generation ability, imposing a constraint that is the opposite of the inherent nature of the large-scale language model 2, namely, "to output a string of characters that is exactly the same as the official notation in the database." So why go to the trouble of suppressing the advantages of large-scale language model 2? The answer is that in the specialized field of trademark practice, slight paraphrasing or summarization by large-scale language model 2 poses the risk of fatally changing the legal meaning. This issue does not exist in general document summarization tasks.
[0047] Next, in the large-scale language model 2, the extracted designated goods or designated services are selected from the candidate group based on the candidate group, input terms, and prompts, and a second candidate group is created (S105). The second group of candidates includes not only the designated goods and designated services, but also their classification, data type, English names of the goods and services, and similar group codes. In other words, the second candidate group is created for each row of the tabular data of the original semantic vector data.
[0048] Next, the control unit 11 receives the selection result (second candidate group) returned from the large-scale language model 2 (S106), and transmits it to the user terminal 1 to display it (S107).
[0049] Here is an example. Figure 4 shows the output to the user terminal 1 when the input term is "judicial scrivener." It can be seen that the output shows "representation in procedures related to registration or deposit," which is the work of a judicial scrivener. Moreover, the output to the user terminal 1 when the input term is "rental" is shown in Figure 5. Even though the target is not limited in this way, various rental business "rental of XX" are picked up.
[0050] According to the designated goods and services selection support system configured as described above, a vector search is performed on a vector database using an input term, multiple designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Japan Patent Office are extracted, the extracted designated goods or designated services and the input term are sent, along with a prompt, to a large-scale language model 2 that can be accessed via a network 100, and the analysis results returned from the large-scale language model 2 are output to a user terminal 1. Therefore, even a person unfamiliar with trademark practice can easily select appropriate designated goods or designated services using everyday language with a low risk of rejection. In other words, even someone unfamiliar with trademark practice can extract designated goods or designated services from the designated goods or designated services actually adopted by the Patent Office, so the indication of the designated goods or designated services will not be unclear, and a reason for refusal (Article 6, Paragraph 1 of the Trademark Act) will not be issued for this reason.
[0051] In particular, because it performs vector searches, even those unfamiliar with trademark practice can arrive at highly relevant candidates for designated goods and services simply by searching using everyday words that come to mind. In addition, the prompt includes analytical rules for identifying the extracted designated goods or designated services that match the meaning indicated by the input term, so that it is possible to derive designated goods or designated services that are deemed appropriate from the vector search candidates in line with the professional criteria of trademark practice and the context.
[0052] Furthermore, the prompt includes output control rules for outputting the identified designated goods or designated services exactly as they are stored in the vector database, without any word-for-word changes, thereby ensuring that the output is an exact match with the designated goods and designated services of the trademark actually adopted by the Patent Office, thereby preventing hallucination.
[0053] Furthermore, the original data used to obtain the semantic vectors is compiled as tabular data, and the semantic vectors are obtained by vectorizing the tabular data row by row. In the vector database, each semantic vector is treated as a single chunk, so it is possible to faithfully extract only relevant items from the normalized list. In other words, hallucination is extremely unlikely to occur, making it possible to select appropriate designated goods or designated services even for input terms entered by someone unfamiliar with trademark practice.
[0054] Furthermore, according to the present invention, the designated goods and designated services of trademarks actually adopted by the Patent Office are in tabular form, and the tabular data includes the English notation of the designated goods and designated services of trademarks actually adopted by the Patent Office, and the semantic vectors are made more accurate by the English notation, thereby improving search accuracy. In other words, using information from English spelling improves embedding, reducing word sense ambiguity, which in turn improves the accuracy of semantic features in vector space and improves the accuracy of search results. This allows for more appropriate selection of designated goods or designated services.
[0055] Furthermore, according to the present invention, the analysis rules are based on the names of the designated goods and designated services and explanatory data that explains the meaning of the names, so that the designated goods and designated services can be more appropriately derived from the candidates searched by vector.
[0056] Second Embodiment Next, a designated product / service selection support system according to a second embodiment of the present invention will be described with reference to Fig. 6. Note that the same parts as those in the first embodiment are denoted by the same reference numerals. The difference between this embodiment and the first embodiment is that the large-scale language model 2 is not used, but the other components are the same as those of the first embodiment.
[0057] That is, the control unit 11 searches the vector database for a vector (S103) and returns the result as is to the user terminal 1 (S108).
[0058] As a result, the output to the user terminal 1 according to this embodiment will contain more noise than the output to the user terminal 1 according to the first embodiment, but the user may check and select from among them himself.
[0059] (Third embodiment) Next, with reference to FIG. 7, a designated product / service selection support system according to a third embodiment of the present invention will be described. When the control unit 11 acquires an input term indicating a product or service input from the user terminal 1 (S101), it sends the input term along with a predetermined prompt to the large-scale language model 2 (S109), and converts the semantic content of the input term from the large-scale language model 2 into an expression suitable for vector search (S110). In other words, the input term is optimized as a search term.
[0060] The control unit 11 then returns the converted input term (search term) (S111), performs a vector search on the vector database using the returned converted input term, and extracts multiple designated goods or designated services that are semantically similar to the converted input term and have actually been adopted by the Patent Office (S112, S113).
[0061] Finally, the control unit 11 transmits the search results (extraction results) of the vector search to the user terminal 1. With the designated goods and services selection support system configured in this way, even a person unfamiliar with trademark practice can easily select appropriate designated goods or designated services with a low risk of rejection using everyday language. This configuration has the advantage of improving the accuracy of the vector search itself.
[0062] In the first embodiment, the analysis rules are based on the names of the designated goods and designated services and explanatory data explaining the meaning of the names, but are not limited to this and may be other analysis rules.
[0063] Furthermore, although the semantic vectors were said to be more accurate by using English notation, the original data for obtaining the semantic vectors does not have to include English notation, as shown in Figure 8. Furthermore, the output to the user terminal 1 only needs to include the designated goods and designated services, and does not necessarily need to include their classification, data type, English names of goods and services, or similar group codes.
[0064] Furthermore, the designated goods and designated services may be output to the user terminal 1 either singly or in plural. Furthermore, it is not necessary to output the information to the user terminal 1 all at once, and questions may be asked of the user to identify the designated goods and designated services.
[0065] Furthermore, semantic vectors are obtained by vectorizing tabular data row by row, and in the vector database, each semantic vector is treated as a single chunk, but this is not limited to this, and it is also possible to vectorize data in other units and treat them as a single chunk.
[0066] In addition, based on the similar group code, which is attribute information of the extracted designated goods or designated services stored in the vector database, other designated goods or designated services that correspond to a higher or lower concept of the designated goods or designated services may be identified. This is based on the information that designated goods or designated services with the same similar group code may include designated goods or designated services of a more general or narrower concept.
[0067] In addition, in the first and third embodiments, prompts and the like are sent to the large-scale language model 2 server, but the large-scale language model 2 server is not limited to a server and also includes forms in which a large-scale language model 2 implemented on a local device is used.
[0068] Although the embedded model is implemented in the management server 10, this is not limited to this and it may be implemented in an external server accessible via the network 100, or it may be implemented in the user terminal 1.
[0069] Furthermore, the entity that generates the semantic vectors may be the management server 10, another processing server, or a local personal computer.
[0070] (Re-ranking / Hybrid Integration / Dynamic K) The control unit 11 may apply re-ranking to the group of candidates obtained by the vector search using a cross-encoder (a cross-encoder is a model that directly estimates a relevance score by simultaneously inputting a query and candidate text) or a bidirectional encoder (a bidirectional encoder is a model that embeds the query and candidate text separately and calculates the similarity using the dot product of the two embeddings, etc.). The control unit 11 may hybrid integrate the vector search results and the results of a word-based search (word-based search refers to a search method based on word similarity such as BM25) using RRF (RRF stands for Reciprocal Rank Fusion, a method of integrating results by taking a weighted sum of the inverses of the ranks) or MMR (MMR stands for Maximal Marginal Relevance, a method of selecting candidates by trading off relevance and diversity). The control unit 11 may dynamically determine the top K to be acquired based on a gap in the similarity distribution or a predetermined threshold value, rather than a fixed value.
[0071] (Methods for achieving strict output / Consistency verification / Ensemble) The output control rules may be realized by at least one of (i) constrained generation based on a JSON schema (JSON schema refers to a constraint specification that defines the key, type, essentiality, etc. of the output) or a formal grammar (formal grammar refers to rules that generate only strings that are allowable based on syntax rules such as BNF), (ii) forcing the output of only candidate IDs by a function call (function call refers to a mechanism that causes the large-scale language model 2 to return only structured arguments such as candidate row IDs), or (iii) vocabulary constraints (vocabulary constraints refer to constraints that prohibit the generation of tokens other than those in the candidate list). The control unit 11 may compare the output of the large-scale language model 2 with the hash value or row ID of the candidate row stored in the vector database, and if there is no exact match, may automatically re-execute the process or switch to a fallback that only presents candidates. The control unit 11 may ensemble the outputs from multiple large-scale language models 2 by self-consistency (self-consistency refers to a method of selecting a stable solution by majority vote of results generated multiple times) or by voting.
[0072] (Preprocessing of input terms / synonym expansion / multilingual processing) The control unit 11 may convert the input terms into query vectors after performing at least some of the following on them: morphological analysis (morphological analysis refers to the process of dividing and annotating words in a sentence into parts of speech, etc.), character normalization (unifying half-width / full-width, old characters / new characters, hiragana / katakana), conjugation normalization, alphanumeric character normalization, unifying punctuation marks and middle dots, absorbing fluctuations in the spelling of foreign words, mutual conversion between romaji and kana, and abbreviation expansion. The control unit 11 may perform a vector search after expanding synonyms with weights using a synonym dictionary (a synonym dictionary is a dictionary that associates synonymous and closely related words, created based on category explanations, etc.). In addition to using English notation, the control unit 11 may also reduce ambiguity in meaning by using multilingual machine translation (multilingual machine translation refers to a method of machine translating input terms into English, etc., and searching in a multilingual embedding space).
[0073] (Data update / version control / duplication elimination) The control unit 11 may extract the difference when an update of the data type is detected, and incrementally update the index by re-embedding only the difference records. The control unit 11 may assign version tags to records according to revisions of classifications, similar group codes, etc., to enable time-specified searches (time-specified searches refer to the function of searching by term groups at a desired reference date and version). The control unit 11 may cluster equivalent expressions and spelling variations, normalize them to a representative spelling, and then register them.
[0074] (Higher, lower, and revised correspondence based on similar group codes) The control unit 11 may present superordinate concepts, subordinate concepts, and related concepts (related concepts refer to synonymous sets within the same code, etc.) as derivation candidates based on the similar group code assigned to the selected candidates. The control unit 11 may allocate equivalent concepts using a code correspondence table (crosswalk) before and after revision (a crosswalk is a table that provides conceptual correspondence between different versions or systems).
[0075] (Generalization to international ID systems) The control unit 11 may generalize the "similar group code" in this specification to the classification, group, and ID system of each country and organization (e.g., the international classification, the ID of each office's acceptable display list, etc.).
[0076] (Dialogue control / Confidence index / Explainability) The control unit 11 may automatically generate and present a disambiguation question (an ambiguity disambiguation question is an additional question to confirm the purpose, target of provision, billing method, etc.) when the similarity is less than a predetermined threshold value or when a conflict between candidates is detected. The control unit 11 may assign each candidate a reliability index (a reliability index is an index that numerically indicates the validity of the output) synthesized from the similarity, re-ranking score, and rule conformance, and may treat candidates that fall below a threshold as pending or for further investigation. The control unit 11 may also add a summary or highlight of the basis fragment such as the category explanation to the output candidate.
[0077] (Performance evaluation and monitoring) The control unit 11 may store evaluation data consisting of pairs of representative query sets (including general terms, industry terms, and misspellings) and correct answer lines, and periodically measure at least one of Precision@k (the rate at which the correct answer is included in the top k results), Hit@1 (the rate at which the correct answer comes in first place), and NDCG (normalized discounted cumulative gain based on the ranking of the correct answer) to monitor quality regression.
[0078] (Implementation form, low latency) The management server 10 and the vector database are not limited to a single-node configuration, and may be implemented as a distributed cluster (a distributed cluster is a configuration that improves scalability and availability through sharding and replication). The control unit 11 may cache the results of frequent queries and update them based on a TTL (TTL means time to live) or invalidation policy.
[0079] (Security and Privacy) The control unit 11 may perform anonymization processing to replace proper nouns corresponding to personal information or trade secrets in the input text with placeholders, thereby minimizing the amount of data transmitted to the large-scale language model 2.
[0080] <Additional Notes> Some or all of the above-described embodiments and examples can be described as, but are not limited to, the following supplementary notes.
[0081] (Appendix 1) A system for supporting the selection of designated goods or designated services, a management server 10 for providing a designated product / service selection support service; a user terminal 1 configured to be able to communicate with the management server 10 over a network 100; The management server 10 a control unit 11 that executes processing related to the designated product / service selection assistance service; a storage unit 12 having a vector database in which the designated goods and designated services of trademarks actually adopted by the Patent Office are pre-stored as semantic vectors; The control unit 11 Acquires an input term indicating a product or service input from the user terminal 1; A vector search is performed on the vector database using the input term to extract multiple designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Patent Office; The designated goods and services selection support system is characterized in that the extracted designated goods or designated services are output to the user terminal 1.
[0082] (Appendix 2) The control unit 11 The designated goods and services selection support system described in Appendix 1 is characterized in that the extracted designated goods or designated services and the input terms are sent to a large-scale language model 2 accessible via the network 100 along with a predetermined prompt, and the analysis results returned from the large-scale language model 2 are output to the user terminal 1.
[0083] (Appendix 3) The analysis rule is: The designated goods and services selection support system described in Supplementary Note 1, characterized in that it includes identifying other designated goods or designated services that correspond to a higher or lower concept of the designated goods or designated services based on the similar group codes assigned to the extracted designated goods or designated services stored in the vector database.
[0084] (Appendix 4) (Query transformation type configuration proposal) A system for supporting the selection of designated goods or designated services, a management server 10 for providing a designated product / service selection support service; a user terminal 1 configured to be able to communicate with the management server 10 over a network 100; The management server 10 a control unit 11 that executes processing related to the designated product / service selection assistance service; a storage unit 12 having a vector database in which the designated goods and designated services of trademarks actually adopted by the Patent Office are pre-stored as semantic vectors; The control unit 11 Acquires an input term indicating a product or service input from the user terminal 1; sending the input terms, together with a predetermined prompt, to a large-scale language model accessible via a network; and returning converted input terms, in which the semantic content of the input terms has been converted into an expression suitable for the vector search, from the large-scale language model; performing a vector search on the vector database using the returned converted input term to extract multiple designated goods or designated services that are semantically similar to the converted input term and have actually been adopted by the Patent Office; The designated goods and services selection support system is characterized in that the extracted designated goods or designated services are output to the user terminal 1.
[0085] (Appendix 5) The designated goods and services selection support system described in either Appendix 1 or 2 is characterized in that the control unit 11 applies re-ranking to the group of candidates obtained by vector search using a cross encoder that simultaneously inputs a query and candidate text to estimate relevance, or a bidirectional encoder that embeds the query and candidate text separately to calculate similarity.
[0086] (Appendix 6) The designated product / service selection support system described in Appendix 5 is characterized in that the control unit 11 integrates vector search results and word-based search results using Reciprocal Rank Fusion (RRF) or Maximal Marginal Relevance (MMR) to obtain hybrid search results.
[0087] (Appendix 7) The designated product / service selection support system described in either Appendix 1 or 2 is characterized in that the control unit 11 dynamically determines the number K of top items to be obtained based on a gap in the similarity distribution or a predetermined threshold value, rather than a fixed value.
[0088] (Appendix 8) The designated goods and services selection support system described in Appendix 2 is characterized in that the output control rules are realized by at least one of constrained generation based on JSON schema or formal grammar, a method that outputs only the identifiers of candidate rows by calling a function, or a vocabulary constraint that prohibits the generation of tokens other than the candidate list.
[0089] (Appendix 9) The designated goods and services selection support system described in Appendix 2 is characterized in that the control unit 11 is equipped with a consistency verification process that compares the output of the large-scale language model 2 with the identifier or hash value of the candidate line stored in the vector database, and automatically re-executes the process in the event of a mismatch, or switches to a fallback that only presents candidates.
[0090] (Appendix 10) The designated goods and services selection support system described in Appendix 2 is characterized in that the control unit 11 ensembles the outputs from multiple large-scale language models 2 using self-consistency or voting based on a majority vote of the results generated multiple times to make a final selection.
[0091] (Appendix 11) The designated goods and services selection support system described in Appendix 1 is characterized in that the control unit 11 performs at least some preprocessing on the input terms, such as normalization of characters and terms, morphological analysis, conjugation normalization, normalization of alphanumeric characters, unification of punctuation marks and midpoints, absorbing fluctuations in the spelling of foreign words, mutual conversion between romaji and kana, and abbreviation expansion, before converting them into a query vector.
[0092] (Appendix 12) The designated product / service selection support system described in Appendix 1 is characterized in that the control unit 11 performs a vector search after weighting and expanding the input terms with synonyms using a synonym dictionary based on category explanations, etc.
[0093] (Appendix 13) The designated product / service selection support system described in Appendix 1 is characterized in that the control unit 11, in addition to using English notation, uses multilingual processing to machine translate input terms into other languages and search them in a multilingual embedded space.
[0094] (Appendix 14) The designated product / service selection support system according to any one of appendices 1 to 4, characterized in that the control unit 11 re-embeds only the differential records when it detects an update to the data type, and incrementally updates the index.
[0095] (Appendix 15) The designated goods and services selection support system according to any one of Supplementary Notes 1 to 4, characterized in that the control unit 11 assigns a version tag to each record according to the revision of the classification or similar group code, thereby enabling a time-specified search based on the version at a specified time.
[0096] (Appendix 16) The designated product / service selection support system according to any one of appendices 1 to 4, characterized in that the control unit 11 clusters a group of records corresponding to equivalent expressions or spelling variations, normalizes them to a representative spelling, and registers them.
[0097] (Appendix 17) The designated goods and services selection support system according to any one of Supplementary Note 1 and 2, characterized in that the control unit 11 presents derived candidates that correspond to a superordinate concept, a subordinate concept, or a synonymous set within the same code based on the similar group code assigned to the selected candidate.
[0098] (Appendix 18) The designated product and service selection support system described in Appendix 17 is characterized in that the control unit 11 performs allocation (crosswalk) of substantially equivalent concepts between different versions based on a code correspondence table before and after the revision.
[0099] (Appendix 19) The designated goods and services selection support system according to any one of Appendices 1 to 4, characterized in that the control unit 11 generalizes the similar group codes in this specification into classification, group and identifier systems adopted by each country or organization.
[0100] (Appendix 20) The designated goods and services selection support system described in either Appendix 1 or 2 is characterized in that the control unit 11 automatically generates and presents additional questions to resolve ambiguity regarding the use, target of provision, or billing form when the similarity is below a predetermined threshold or when a conflict between candidates is detected.
[0101] (Appendix 21) The control unit 11 assigns a reliability index to each candidate that is synthesized from the similarity, re-ranking score, and rule conformance, and treats candidates whose index is below a threshold as pending or for further investigation, as described in either Appendix 1 or 2.
[0102] (Appendix 22) The designated product / service selection support system according to either Appendix 1 or 2, characterized in that the control unit 11 includes a summary or highlight of the basis fragment based on the category explanation, etc., in addition to the output candidates.
[0103] (Appendix 23) The control unit 11 uses evaluation data consisting of pairs of representative query sets and correct answer lines, and periodically measures at least one of Precision@k, Hit@1, or Normalized Discounted Cumulative Gain (NDCG) to monitor quality, as described in any of Appendices 1 to 4.
[0104] (Appendix 24) A designated product / service selection support system according to any one of appendices 1 to 4, characterized in that the management server 10 and vector database are implemented in a distributed cluster configuration including sharding and replication.
[0105] (Appendix 25) The designated product / service selection support system according to any one of appendices 1 to 4, characterized in that the control unit 11 caches the results of frequently occurring queries and updates the cache by applying an expiration date or invalidation policy.
[0106] (Appendix 26) The designated goods and services selection support system described in either Appendix 1 or 2 is characterized in that the control unit 11 performs an anonymization process to replace proper nouns in the input text that correspond to personal information or trade secrets with placeholders, thereby minimizing the amount of data transmitted to the large-scale language model 2.
[0107] (Appendix 27) The designated goods and services selection support system described in Appendix 2 is characterized in that the large-scale language model 2 is not limited to being provided by a server on the cloud, but can be switched to a form in which inference is performed using a large-scale language model 2 implemented on an on-premise or local device.
[0108] (Appendix 28) A program that allows a computer that can access a vector database in which the designated goods and designated services of trademarks actually adopted by the Patent Office are pre-stored as semantic vectors to function as a device that supports the selection of designated goods or designated services, The computer, an acquisition step of acquiring input terms indicating products or services input from the user terminal 1; an extraction step of performing a vector search on the vector database using the input term to extract multiple designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Patent Office; an output step of outputting the extracted designated goods or designated services to the user terminal 1; A designated goods and services selection support program characterized by executing the above.
[0109] (Appendix 29) A designated product / service selection support program as described in Appendix 28, characterized in that the program further executes a re-ranking step of re-ranking the group of candidates obtained by the extraction step using a cross encoder or a bidirectional encoder.
[0110] (Appendix 30) A designated goods and services selection support program as described in Appendix 28, further comprising: a word-based search step for executing a word-based search; and an integration step for integrating the results of the extraction step and the word-based search step by Reciprocal Rank Fusion or Maximal Marginal Relevance.
[0111] (Appendix 31) A program characterized in that the designated product / service selection support program described in Appendix 28 further executes a number determination step of determining the top number K of items to be obtained based on a gap in the similarity distribution or a predetermined threshold value.
[0112] (Appendix 32) A designated product / service selection support program as described in Appendix 28, further comprising: a transmission step of transmitting the group of candidates and the input terms to a large-scale language model 2 together with a predetermined prompt; and a constraint output step of forcing the generated output of the large-scale language model 2 to output only the identifiers of the candidate lines by using a JSON schema, a formal grammar, or a function call.
[0113] (Appendix 33) 33. The program according to claim 32, further comprising a consistency verification step of comparing the output of the large-scale language model 2 with the identifier or hash value of a candidate line stored in a vector database, and switching to automatic re-execution or fallback of only presenting candidates when there is a mismatch.
[0114] (Appendix 34) 33. The program according to claim 32, further comprising an ensemble step of making a final selection by self-consistency or voting on the outputs of a plurality of large-scale language models 2.
[0115] (Appendix 35) A designated goods and services selection support program as described in Appendix 28, characterized in that the program further executes a preprocessing step for input terms, which performs at least some of the following: normalization of characters and terms, morphological analysis, conjugation normalization, normalization of alphanumeric characters, unification of punctuation marks and midpoints, absorption of fluctuations in the spelling of foreign words, mutual conversion between romaji and kana, and expansion of abbreviations.
[0116] (Appendix 36) A program according to claim 28, further comprising a synonym expansion step of expanding input terms with weights using a synonym dictionary based on category descriptions, etc., and then performing a vector search.
[0117] (Appendix 37) 29. The program according to claim 28, further comprising a multilingual processing step of machine-translating input terms into other languages and searching them in a multilingual embedding space, in addition to using English notation.
[0118] (Appendix 38) The program according to claim 28, further comprising an incremental update step of re-embedding only differential records and incrementally updating the index when an update of the data type is detected.
[0119] (Appendix 39) A program according to Appendix 28, characterized in that it further executes a time-specified search step of assigning a version tag to each record in accordance with the revision of the classification or similar group code, and conducting a search based on the version at the specified time.
[0120] (Appendix 40) A program as described in Appendix 28, characterized in that it further executes a normalization registration step of clustering records corresponding to equivalent expressions or spelling variations and normalizing them to a representative spelling.
[0121] (Appendix 41) The program according to Appendix 28, further comprising a hierarchical expansion step of presenting derived candidates corresponding to superordinate concepts, subordinate concepts, or synonymous sets within the same code, based on the similar group code assigned to the selected candidates.
[0122] (Appendix 42) A program as described in Appendix 41, characterized in that it further executes a crosswalk step that allocates substantially equivalent concepts between different versions based on code correspondence tables before and after the revision.
[0123] (Appendix 43) A program as set forth in Appendix 28, further comprising an identifier generalization step for generalizing the similar group codes in this specification into classification, group, and identifier systems adopted by each country or organization.
[0124] (Appendix 44) The program described in Appendix 28 further comprises a question generation step of automatically generating and presenting an additional question to resolve ambiguity regarding the use, target of provision, or billing method when the similarity is less than a predetermined threshold or when a conflict between candidates is detected.
[0125] (Appendix 45) The program according to claim 28, further comprising a reliability control step of calculating a reliability index for each candidate that is synthesized from the similarity, re-ranking score, and rule conformance, and treating candidates whose index is below a threshold as pending or for further investigation.
[0126] (Appendix 46) A program as described in Appendix 28, characterized in that the program further executes a basis presentation step of displaying a summary or highlight of the basis fragment based on the category explanation, etc., alongside the output candidate.
[0127] (Appendix 47) 29. The program according to claim 28, further comprising an evaluation monitoring step for periodically measuring at least one of Precision@k, Hit@1, or Normalized Discounted Cumulative Gain (NDCG) to monitor quality using evaluation data consisting of pairs of a representative set of queries and correct answers.
[0128] (Appendix 48) 29. The program according to claim 28, further comprising a cache management step of caching results of frequently occurring queries and updating the cache by applying an expiration date or invalidation policy to the cache.
[0129] (Appendix 49) 29. The program of claim 28, further comprising an anonymization step of replacing proper nouns in the input text that correspond to personal information or trade secrets with placeholders to minimize the amount of data transmitted to the large-scale language model 2.
[0130] (Appendix 50) 29. The program according to claim 28, further comprising a destination switching step of switching the inference destination of the large-scale language model 2 between a server on the cloud and a model on an on-premise or local device. [Explanation of symbols]
[0131] 1. User terminal 2 Large-scale language models 10 Management Server 11 Control section 12 Storage section 100 Network
Claims
1. A system for supporting the selection of designated goods or designated services, a management server for providing a designated goods and services selection support service; a user terminal configured to be able to communicate with the management server over a network, The management server a control unit that executes processing related to the designated product / service selection assistance service; a storage unit having a vector database in which semantic vectors based on the designated goods and designated services of trademarks actually adopted by the Patent Office are stored in advance; The control unit Acquire an input term indicating a product or service input from the user terminal; A vector search is performed on the vector database using the input term, and a plurality of designated goods or designated services that are semantically similar to the input term and have actually been adopted by the Patent Office are extracted; sending the extracted designated goods or designated services and the input terms, together with a prompt, to a large-scale language model accessible via a network; Execute a process of outputting the analysis result returned from the large-scale language model to the user terminal; The prompt may include: (a) an analysis rule for identifying, from among the extracted designated goods or designated services, those that match the meaning indicated by the input term; (b) an output control rule for outputting the identified designated goods or designated services in the same format as stored in the vector database, without any changes to the description.
2. The original data for obtaining the semantic vector is compiled as tabular data, the semantic vector is obtained by vectorizing the tabular data row by row, 2. The designated product / service selection support system according to claim 1, wherein each of said semantic vectors is treated as one chunk in said vector database.
3. The original data for obtaining the semantic vector is compiled as tabular data, The tabular data includes the English names of the designated goods and designated services of the trademarks actually adopted by the Patent Office, 2. The designated goods and services selection support system according to claim 1, wherein the semantic vector is made more accurate by using the English notation.
4. The analysis rule is: The names of the designated goods and services; 4. The designated goods / services selection support system according to claim 1, further comprising explanation data explaining the meaning of the name.